PhD Chapter 3
Results 3/3
This series of files compile all analyses done during Chapter 3:
- Section 1 presents the calculation of the indices of exposure.
- Section 2 presents variable exploration and regressions results.
- Section 3 presents species distribution models.
All analyses have been done with R 4.0.2.
Click on the table of contents in the left margin to assess a specific analysis.
Click on a figure to zoom it
Sources of activity considered for the analyses:
- aquaculture: mussel farm (AquaInf)
- city: general diffusive influence, wharves (CityInf, CityWha)
- industry: general diffusive influence, wharves (Indu, InduWha)
- sediment dredging: collection zones, dumping zones (DredColl, DredDump)
- commercial shipping: mooring sites, traffic routes (ShipMoor, ShipTraf)
- sewers: rainwater drains, wastewater drains (SewRain, SewWast)
Fisheries data considered for the analyses (expressed as number of fishing events or kilograms of collected individuals for each gear):
| Gear | Code | Years | Events | Species |
|---|---|---|---|---|
| Dredge | FishDred | 2010-2014 | 21 | Mactromeris polynyma |
| Net | FishNet | 2010 | 5 | Clupea harengus, Gadus morhua |
| Trap | FishTrap | 2010-2015 | 1061 | Buccinum sp., Cancer irroratus, Chionoecetes opilio, Homarus americanus |
| Bottom-trawl | FishTraw | 2013-2014 | 2 | Pandalus borealis |
1. Methodology
The aim of this section is to predict the structure of benthic communities based on the values of environmental variables.
We used abiotic parameters and indices of human exposure indices (calculated in Section 1) as predictors. We tested different methods: GLMs, GAMs, Random Forest and HMSC. Each method has been developed in dedicated scripts, whose final objects were imported here to present results and trends.
For each method, results are presented with a table regrouping McFadden’s or Tjur’s pseudo-R2, validation ratios and variables coefficients, and with maps displaying the probability of presence of each taxon for which the pseudo-R2 is higher than 0.20 (and different than 1). The raster presents results of the SDM (grey: low probability, dark blue: high). Stations are either plotted with colors (green = taxa present, red = taxon absent) or with circles (wider circcles = higher taxon density).
2. Models
2.1. Generalized Linear Models
Diagnostics for each model can be found here.
2.1.1. Presence/absence data
We considered presence/absence data with a binomial distribution.
Abiotic parameters
Presence probability of significative taxa:
Prediction of specific richness based on this model:
Exposure indices
Presence probability of significative taxa:
Prediction of specific richness based on this model:
2.1.2. Density data
⚠️ To be added … or not
2.2. Hierarchical Models of Species Communities
This section uses methodology and tools from Ovaskainen et al., with the direct help of Guillaume Blanchet.
First, we will compute models using the 108 stations with abiotic variables or exposure indices as predictors. 85 % of the stations (92) will act as training data, and the rest (16) will be used to validate the outputs. Second, these models will be used to predict taxa richness and distribution in the entire study area using predictor rasters.
We initiate the HMSC model with the chosen data:
- presence/absence or density for dependant variable
- exposure indices or abiotic variables for predictors
Priors and model parameters are set in the hmsc() function.
HMSC_PA <- hmsc(data, param, prior, family = "probit", niter = 100000, nburn = 1000, thin = 100)
HMSC_density <- hmsc(data, param, prior, family = "overPoisson", niter = 100000, nburn = 1000, thin = 100)Here are the outcomes and diagnostics to evaluate each model’s quality (presented for each species seperately or averaged).
Diagnostics for each model can be found here.
2.2.1. Presence/absence data
We considered presence/absence data with a probit distribution.
Abiotic parameters
Mean of the predictor coefficients estimated by the HMSC model:
95 % confidence interval of the predictor coefficients estimated by the HMSC model:
Predictive power of the HMSC model:
Variance partitioning:
Presence probability of significative taxa:
Prediction of specific richness:
Exposure indices
Mean of the predictor coefficients estimated by the HMSC model:
95 % confidence interval of the predictor coefficients estimated by the HMSC model:
Predictive power of the HMSC model:
Variance partitioning:
Presence probability of significative taxa:
Prediction of specific richness:
2.2.2. Density data
⚠️ To be added … or not
2.3. Generalized Additive Models
2.3.1. Presence/absence data
⚠️ To be added … or not
2.3.2. Density data
⚠️ To be added … or not
2.4. Random Forest Algorithms
2.4.1. Presence/absence data
⚠️ To be added … or not
2.4.2. Density data
⚠️ To be added … or not